In-furnace 3D temperature field reconstruction method and system

Abstract

The invention discloses an in-furnace 3D temperature field reconstruction method and system, and relates to the technical field of blast furnace temperature detection.The method comprises the steps that a reference database of a temperature field is established; acquiring new working condition data, selecting reference working condition data most similar to the new working condition data based on the reference database, and acquiring a projection temperature point and a corresponding temperature change value according to the reference working condition data and the new working condition data; calculating sagittal view and coronal view center slice images of the new working condition data according to the projection temperature points and the corresponding temperature change values in combination with an inverse distance weighted interpolation method; the combined action of thermodynamic factors in the hearth is simulated through sagittal view and coronal view center slice images of the new working condition data by ray casting, a transection image of the new working condition data is reconstructed in combination with an adversarial neural network, and a reconstructed 3D hearth temperature field is obtained. According to the invention, rapid establishment of the 3D temperature field of the hearth is realized through the direct temperature measurement point, the water-cooled wall temperature measurement point and the boundary condition.

Description

Technical Field

[0001] The present invention relates to the technical field of blast furnace temperature detection, and more particularly to a method and system for reconstructing a 3D temperature field in a furnace. Background Art

[0002] At present, coal-fired power generation is still the main means of power generation in China. Therefore, it is very important to maintain the stability of the flame during the combustion process of coal-fired boilers. However, the combustion process of coal-fired boilers is complicated, and the extreme environment in the furnace makes temperature measurement difficult. Traditional boiler combustion detection devices generally use industrial television monitoring methods, namely furnace safety monitoring systems, which judge the combustion status of the boiler by the color and brightness of the flame to ensure that the combustion flame does not go out. It can continuously and closely monitor the parameters and status of the combustion system under various operating conditions such as normal operation and start-up and shutdown of the boiler, continuously perform logical judgments and calculations, and issue action instructions when necessary, so that the combustion equipment can complete the operation according to a reasonable procedure to ensure the safety of boiler combustion. Traditional temperature measurement equipment can only monitor several measurement points in the furnace, but cannot monitor the overall temperature field of the furnace. Due to the lack of an effective overall temperature field monitoring method, the timely and effective adjustment of the combustion process is seriously hindered, resulting in the inability of traditional control strategies to effectively reduce fuel consumption and meet the needs of flexible operation of the combustion system. Therefore, it is crucial to quickly and accurately establish the 3D temperature field of the furnace for the operation and control of the boiler. Summary of the invention

[0003] In view of this, the present invention provides a method and system for reconstructing the 3D temperature field in a furnace, which realizes the rapid establishment of the 3D temperature field in the furnace through direct temperature measurement points, water-cooled wall temperature measurement points and boundary conditions.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] A method for reconstructing a 3D temperature field in a furnace, comprising:

[0006] Establishing a reference database of the temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images;

[0007] Acquire new working condition data, select working condition data that is most similar to the new working condition data based on the reference database as reference working condition data, acquire projected temperature points and corresponding temperature change values ​​based on the reference working condition data and the new working condition data; acquire a temperature difference map based on the projected temperature points and the corresponding temperature change values ​​in combination with an inverse distance weighted interpolation method, and calculate sagittal and coronal central slice images of the new working condition data based on the temperature difference map;

[0008] Through the sagittal and coronal central slice images of the new operating condition data, the combined effect of thermodynamic factors in the furnace is simulated by ray projection, and all cross-sectional images of the new operating condition data are reconstructed in combination with the pre-trained adversarial neural network to obtain the reconstructed 3D furnace temperature field.

[0009] Preferably, the selection of reference operating condition data that is most similar to the new operating condition data based on the reference database specifically includes: finding the load corresponding interval of the new operating condition data, and finding the operating condition that is most similar to the new operating condition data and the operating condition data in the interval through KNN clustering as the reference operating condition data; the KNN clustering features select the direct temperature measurement point data and boundary condition data.

[0010] Preferably, the step of acquiring the projected temperature points and the corresponding temperature change values ​​according to the reference operating condition data and the new operating condition data specifically includes: the projected temperature points include water-cooled wall temperature points and direct temperature measurement points; is the central slice image I in the sagittal position s The water-cooled wall temperature point projected on the top, The central slice image I in the coronal position c The temperature difference between the new working condition data and the reference working condition data at the water-cooled wall temperature point is ΔT. The projected temperature value should have a corresponding change. The change satisfies the nonlinear transformation relationship. The formula is as follows:

[0011] ΔT c =F1(ΔT)

[0012]

[0013]

[0014] Where, ΔT c Indicates that the direct temperature measurement point is in the central slice image I of the coronal position c Temperature change value of the upper projection point, ΔT w_s ,ΔT w_c Respectively represent the water-cooled wall temperature measurement point in the sagittal and coronal center slice images I s , I c The temperature change value of the upper projection point, F1, G2, F2 represents the nonlinear transformation relationship.

[0015] Preferably, the calculation of the sagittal and coronal central slice images of the new working condition data based on the temperature difference map specifically includes: using the inverse distance weighted interpolation method to propagate the temperature change value to I s , I c Get the temperature difference map on the whole image;

[0016] The sparse projection set {ΔTp |ΔT c ,ΔT w_s ,ΔT w_c} as the observed value, I s , I c The coordinate points to be filled in are regarded as points to be estimated, and the temperatures to be estimated are regarded as values ​​to be estimated;

[0017] Calculate the weight coefficients of the M observation points closest to the point to be estimated for the point to be estimated, as follows:

[0018]

[0019]

[0020]

[0021]

[0022] Among them, W idw is the weight coefficient, (u0,v0) is the spatial coordinate of the point to be estimated, (u j ,v j ) is the spatial coordinate of the observation point, d j is the Euclidean distance from the estimated point j to the observation point, is the value to be estimated, q is a constant, T j are the j closest observation points;

[0023] The multi-layer inverse distance weighted interpolation method is used to interpolate the projection points to obtain the intermediate point set, and then the intermediate point set is interpolated again to obtain I s , I c All the points in the above figure are interpolated to obtain the temperature difference between the sagittal and coronal central slices. d_s , I d_c , combined with the sagittal and coronal central slice images corresponding to the reference working condition data I r_s , I r_c , get I s , I c , as shown below:

[0024] I s , I c =I r_s , I r_c +I d_s , I d_c .

[0025] Preferably, reconstructing all the transverse images of the new working condition data specifically comprises: s , I cPerform feature extraction, extract the position of slice i from the 3D position, and convert the feature information, slice position, and sagittal observation position V s and coronal observation position V c Perform ray projection, combine the feature information with the 3D position through feature space projection, and then decode the feature through the decoder to generate the image of the cross-sectional slice i

[0026] Preferably, the encoder is used to s , I c The feature extraction specifically includes: the feature encoder uses resnet to extract features, and uses the resnet to extract I s , I c Feature extraction is performed separately to obtain:

[0027] Res(I s , I c )=(f s , f c )

[0028] Among them, f s , f c Indicates that from I s , I c Extracted intermediate level feature information.

[0029] Preferably, the combining of the feature information with the 3D position through feature space projection specifically comprises: for a slice i in the 3D position, for each 3D point p g_i =(x g_i ,y g_i ,z g_i )To I s , I c projection:

[0030] (u i ,v i ) T =[R|t](x g_i ,y g_i ,z g_i ) T ;

[0031] Where R and t represent the rotation matrix and translation vector respectively. s , I c The extracted intermediate level features f s , f c In association, the method B of sampling bilinear interpolation of the association method is expressed as:

[0032] f i =C(B(f s ,us_i ,v s_i ),B(f c ,u c_i ,v c_i ))

[0033] Where C represents the stacking operation, f i That is to reconstruct the features of slice i, (u s_i ,v s_i ) is the sagittal center slice projection point, (u c_i ,v c_i ) is the projection point of the coronal center slice.

[0034] Preferably, the feature decoding by a decoder specifically includes:

[0035]

[0036] in, is the image of the cross-section slice i, and D is the decoder.

[0037] A furnace 3D temperature field reconstruction system, comprising:

[0038] A database establishment module is used to establish a reference database of the temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images;

[0039] a slice image calculation module, which obtains new working condition data, selects working condition data that is most similar to the new working condition data based on the reference database as reference working condition data, obtains projected temperature points and corresponding temperature change values ​​based on the reference working condition data and the new working condition data; obtains a temperature difference map based on the projected temperature points and the corresponding temperature change values ​​in combination with an inverse distance weighted interpolation method, and calculates sagittal and coronal central slice images of the new working condition data based on the temperature difference map;

[0040] The temperature field reconstruction module uses the sagittal and coronal central slice images of the new operating condition data to simulate the combined effect of thermodynamic factors in the furnace through ray projection, and combines the pre-trained adversarial neural network to reconstruct all cross-sectional images of the new operating condition data to obtain a reconstructed 3D furnace temperature field.

[0041] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a method and system for reconstructing the 3D temperature field in a furnace, which fully considers the correction effect of the water-cooled wall temperature measuring points and the direct temperature measuring points on the temperature field, so that the generated coronal and sagittal central slice images are closer to the actual temperature distribution in the furnace, and proposes a method for mapping 2D features to 3D space through a ray projection model, which can restore the entire three-dimensional temperature field from the coronal and sagittal central slice images. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 The overall framework diagram provided for the present invention;

[0044] Figure 2 A two-stage technical roadmap provided for the present invention;

[0045] Figure 3 A schematic cross-sectional view of a front and rear wall boiler provided by the present invention;

[0046] Figure 4 A grid diagram of a 300MW boiler provided by the present invention in Fluent software;

[0047] Figure 5 Orthogonal projection diagram of the water-cooled wall and direct temperature measurement point provided by the present invention;

[0048] Figure 6 The second-stage network structure diagram provided by the present invention;

[0049] Figure 7 The network output result diagram provided by the present invention;

[0050] Figure 8 A slice definition diagram provided by the present invention;

[0051] Fig. 9 The result denoised image provided by the present invention;

[0052] Fig.10 This is a schematic diagram of the temperature field results provided by the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The embodiment of the present invention discloses a method for reconstructing a 3D temperature field in a furnace, comprising:

[0055] Establish a reference database of temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images;

[0056] Acquire new working condition data, select working condition data that is most similar to the new working condition data based on the reference database as reference working condition data, and acquire the projection temperature points and corresponding temperature change values ​​based on the reference working condition data and the new working condition data; acquire a temperature difference map based on the projection temperature points and corresponding temperature change values ​​combined with the inverse distance weighted interpolation method, and calculate the sagittal and coronal central slice images of the new working condition data based on the temperature difference map;

[0057] Through the sagittal and coronal central slice images of the new operating condition data, the combined effect of thermodynamic factors in the furnace is simulated by ray projection, and all cross-sectional images of the new operating condition data are reconstructed in combination with the pre-trained adversarial neural network to obtain the reconstructed 3D furnace temperature field.

[0058] In a specific embodiment, the newly-incoming operating condition data refers to newly-incoming direct temperature measurement point data, newly-incoming water-cooled wall temperature measurement point data, and newly-incoming boundary condition data.

[0059] In Example 1, a 3D temperature field modeling is performed on a typical working condition by using a CFD (computational fluid dynamics) method to establish a reference database of the temperature field;

[0060] Based on direct temperature measurement points {T1...T 40}、Water-cooled wall temperature measurement point {T w1 ...T w559}, boundary conditions {C1...C 30}, the complete N×N×N temperature field can be reconstructed. In this embodiment, a 128×128×128 temperature field is taken as an example to explain the details of the proposed method. The method shows potential in field applications and has the effect of fast reconstruction. The reconstructed temperature field takes less than 1 minute on a 3090 graphics card. Compared with the results of the CFD method slice by slice, the average PSNR is 29, the average SSIM is 0.96, and the average Lpips is 0.02.

[0061] In the first stage, the sagittal and coronal central slice images of the 3D temperature field were calculated based on KNN clustering, multi-layer perceptron, inverse weighted distance interpolation and other methods. s , I c , and use direct temperature measurement data and water wall data to make I s , I c The temperature distribution restoration is closer to the real result, and the subtraction operation of the reference working condition data and the new working condition data is used to eliminate the uncertainty affecting the temperature restoration.

[0062] The second stage of the proposed method combines the physics-driven model of ray casting with a generative neural network to learn the normalized 3D position and I s , I c The relationship between features is analyzed, the features are spatially mapped, and the mapped features are decoded to obtain cross-sectional slice images, thereby generating a complete 128×128×128 furnace temperature field.

[0063] Figure 1 For the overall framework of the present invention, in the training stage, a temperature field database is first established to obtain the 3D temperature field of typical working conditions. This database will be divided into two parts, one as a reference database and the other for training the neural network in the two-stage method of design, such as Figure 2 As shown, during the testing phase, the newly added direct temperature measurement points, water-cooled wall temperature measurement points and boundary conditions are input into the two-stage method to generate a complete 3D temperature field in the furnace.

[0064] This embodiment uses a front and rear wall hedge coal-fired boiler, such as Figure 3 As shown, the height is 55.56m, the width is 20.67m, and the depth is 15.73m. There are 4 layers of burners, 5 on each layer of the front wall and the rear wall, totaling 40 burners. The direct temperature measuring points are arranged according to the number of burners, starting from a height of 28.22m to an end at a height of 43.46m, with a row arranged every 5.08m, for a total of four rows. Each row starts from 3.17m to 19.03m, and a temperature measuring point is arranged at intervals of 3.965m. There are 5 temperature measuring points in each row, 20 on the front wall, 20 on the rear wall, for a total of 40.

[0065] In a specific embodiment, selecting reference operating condition data that is most similar to the new operating condition data based on the reference database specifically includes: finding the load corresponding interval of the new operating condition data, and finding the most similar operating condition between the new operating condition data and the 10 groups of operating condition data in the interval as the reference operating condition data through KNN clustering, as shown in Table 1 in this embodiment; the KNN clustering features selected are direct temperature measurement point data and boundary condition data.

[0066] Table 1 Schematic diagram of boundary conditions

[0067]

[0068]

[0069] The rated load of the boiler is 300MW, and the working load range is concentrated between 150MW and 300MW. The range is evenly divided into 5 intervals, each interval is 30MW, and 10 groups of data are selected for a single interval. The SpaceClaim component built into the numerical simulation software ANSYSFluent is used to construct a three-dimensional model of the furnace. In order to reduce costs and ensure the accuracy of the calculation results, the boiler model is appropriately simplified, retaining only the main part of the furnace, including the burner area (including OverFire Air and Primary Air), hopper, screen superheater, final superheater, high-temperature reheater and other areas. The ICEM component in ANSYS Fluent is used to perform structured meshing on the three-dimensional furnace model, and the mesh is refined in the burner area. There are a total of 867,046 meshes, such as Figure 4 As shown. In the numerical simulation process carried out in ANSYS Fluent, in order to conform to the physical characteristics of the actual combustion process, we adopted the law of conservation of mass, law of conservation of momentum, and law of conservation of energy in the energy equation. The Realizable-k-ε model takes into account the turbulent pulsation velocity and pulsation length at the same time, which is more in line with the actual situation of boiler combustion. We choose it as the turbulence model. Radiation heat transfer is often involved in the combustion system. The P1 model has the characteristics of easy processing of complex geometric figures and low CPU resource consumption. We choose it as the radiation model. We choose the component transport model to simulate the chemical reaction between the various material components in the boiler combustion process. In order to solve the coupling problem of chemical reaction and turbulence, we choose the Eddy-Dissipation model. The size, density, and volatile content of the pulverized coal particles in the combustion process will affect the accuracy of the numerical simulation results. We choose the Two-Competing-Rates model to simulate the competitive reaction between the volatilization of pulverized coal and solid carbon. The boundary conditions set in the numerical simulation process are shown in Table 1.

[0070] The reconstructed result is compared with the temperature of the water-cooled wall to check whether the temperature distribution of the reconstructed temperature field is consistent with the temperature distribution of the water-cooled wall. If the distribution is inconsistent, the boundary conditions are repeatedly adjusted according to the guidance of multiple experts in on-site boiler operation and maintenance, and the data is sent to the Fluent software for calculation again until the distribution is consistent. In this embodiment, 200 sets of data are accurately reconstructed by Fluent software. During use, 50 sets are used as reference data sets, 120 sets are used as training data for the first and second stages of multi-layer perceptrons and generative adversarial neural networks, and 30 sets are used as corresponding test data.

[0071] The first stage calculates the sagittal and coronal central slice images I corresponding to the new working condition data C. s , I c , first estimate I s , I c The reason is that these two slices can include the projection values ​​of all water-cooled wall data and the projection values ​​of direct temperature measurement points. With the help of water-cooled wall distribution, I s , I c The temperature distribution restoration is easier to approach the real result. s , I c Reconstruction process as Figure 2 As shown on the left, first find the working condition C that is most similar to C from the reference database r , based on C and C r Direct temperature measurement data and water wall data calculation I s , I c and I r_s , I r_c The difference between the images I d_s , I d_c , where I r_s , I r_c It is C r The sagittal and coronal central slice images of the state are combined with I r_s , I r_c Calculate I s , I c .

[0072] Find C r After that, estimate the difference map I d_s , I d_c , there is no direct estimate of I s , I c The reason is that there are many random factors affecting the boiler temperature field, and the uncertainty is relatively large. The deviation of direct estimation will be large. Using the difference map can offset a lot of uncertainty through subtraction, and the estimation accuracy is higher. r The temperature difference of the water-cooled wall temperature measuring point is {ΔT w |T w1 -T r_w1 ,...,T w559 -T r_w559}, where {T w1 ...T w559} and {T r_w1 ...T r_w559} represent C and C respectively r The water wall temperature, Indicates ΔT w Can be found in I s There are 262 water-cooled wall temperature points projected on the top. Indicates ΔT w Can be found in I c There are 297 water-cooled wall temperature points projected on the top. The temperature difference between the current working condition and the reference working condition is {ΔTT1-T r_1 ,...,T 40 -T r_40}, where {T1...T 40} and {T r_1 ...T r_40} distribution indicates that C and C r The temperature of the direct temperature measurement point can only be projected to I due to the front and rear wall boilers. c , cannot be projected onto I s . is the central slice image I in the sagittal position s The water-cooled wall temperature point projected on the top, The central slice image I in the coronal position c The temperature difference between the water-cooled wall temperature point projected on the upper side and the temperature point directly measured by the new working condition data and the reference working condition data is ΔT. and ΔT to I s , I c Perform orthographic projection, such as Figure 5 As shown, the temperature value after projection should have a corresponding change, and the change satisfies the nonlinear transformation relationship (that is, first obtain the nonlinear transformation relationship between the actual temperature value and the projected temperature value through the multi-layer perceptron, and then obtain the corresponding projected temperature value according to the nonlinear transformation relationship). The formula is as follows:

[0073] ΔT c =F1(ΔT)

[0074]

[0075]

[0076] Where, ΔT c Indicates that the direct temperature measurement point is in the central slice image I of the coronal position c Temperature change value of the upper projection point, ΔT w_s ,ΔT w_c Respectively represent the water-cooled wall temperature measurement point in the sagittal and coronal center slice images I s , I cThe temperature change value of the projection point on the left, F1, G2, F2 represent the nonlinear transformation relationship, which is fitted by a multi-layer perceptron. The perceptron includes an input layer, three hidden layers, and an output layer. The connection relationship is that the input layer is connected to the three hidden layers in sequence, and then connected to the output layer. For each formula in the above formula, a multi-layer perceptron model needs to be trained separately, for a total of three multi-layer perceptron models. After the three multi-layer perceptron models, the corresponding positions in the difference map are filled with the corresponding temperature difference values, and then the temperature difference values ​​are broadcast to I using the inverse distance weighted interpolation method. s , I c on the entire image.

[0077] In one specific embodiment, the temperature change value is propagated to I using the inverse distance weighted interpolation method. s , I c The temperature difference map is obtained on the entire image of the temperature field. The principle of applying the inverse distance weighted interpolation method to temperature field reconstruction is that as the distance increases, the influence of the value of any temperature measurement point on the values ​​of the surrounding temperature measurement points gradually decreases. This method converts the sparse projection set {ΔT p |ΔT c ,ΔT w_s ,ΔT w_c} as the observed value. s , I c The coordinate points to be filled in are regarded as the points to be estimated, and the temperature to be estimated is regarded as the estimated value. When determining the estimated value, it is assumed that the M observation points closest to the estimated value have an influence on it. These M observation points have an influence on the estimated point inversely proportional to the distance. Therefore, the estimated points that are closer to each other will be given greater weights, and the sum of all weights is equal to 1.

[0078] The sparse projection set {ΔT p |ΔT c ,ΔT w_s ,ΔT w_c} as the observed value, I s , I c The coordinate points to be filled in are regarded as points to be estimated, and the temperatures to be estimated are regarded as values ​​to be estimated;

[0079] Calculate the weight coefficient of the M observation points closest to the point to be estimated. In the present invention, M is 11, and the weight coefficient W idw The calculation of is the key to the IDW interpolation algorithm:

[0080]

[0081]

[0082]

[0083]

[0084] Among them, W idw The weight coefficient is determined by the power of the inverse of the distance, which means that as the distance increases, the weight decreases proportionally, and the rate of weight reduction depends on q, where q = 2. The Euclidean distance is used to measure the distance from the observation point to all the predetermined estimation points. (u0, v0) is the spatial coordinate of the point to be estimated, (u j ,v j ) is the spatial coordinate of the observation point, d j is the Euclidean distance from the estimated point j to the observation point, is the value to be estimated, q is a constant, T j The closest j observation points are substituted into other formulas to obtain;

[0085]

[0086] The multi-layer inverse distance weighted interpolation method is used to interpolate the projection points to obtain the intermediate point set, and then the intermediate point set is interpolated again to obtain I s , I c All the points in the above figure are interpolated to obtain the temperature difference between the sagittal and coronal central slices. d_s , I d_c , combined with the sagittal and coronal central slice images corresponding to the reference working condition data I r_s , I r_c , get I s , I c , as shown below:

[0087] I s , I c =I r_s , I r_c +I d_s , I d_c .

[0088] Among them, the intermediate point set includes 317 measurement points in the coronal plane (277 water-cooled walls + 20 direct measurement points on the front wall, and 20 on the back wall), of which 1188 points were obtained by the first interpolation and 21834 points were obtained by the second interpolation, and 282 measurement points in the sagittal plane (all water-cooled walls), of which 842 points were obtained by the first interpolation and 12507 points were obtained by the second interpolation.

[0089] The second stage is to calculate the I s , I c , reconstruct all transverse images The technical route at this stage is as follows Figure 2 As shown on the right. This embodiment uses an adversarial neural network with an encoder and a decoder. The network structure is as follows Figure 6 As shown, the encoder iss , I c Perform feature extraction, extract the position of slice i from a set of normalized 3D positions, and combine the features, slice positions, and sagittal observation positions V s and coronal observation position V c Perform raycasting to combine features with 3D positions, and then decode the features through a set of decoders to generate an image of the cross-sectional slice i The second stage uses adversarial neural network training. The trained neural network generates all cross-sectional slices in parallel through GPU in the actual reasoning stage.

[0090] In a specific embodiment, reconstructing all transverse images of the new working condition data specifically includes: s , I c Perform feature extraction, extract the position of slice i from the 3D position, and convert the feature information, slice position, and sagittal observation position V s and coronal observation position V c Perform ray projection, combine the feature information with the 3D position through feature space projection, and then decode the feature through the decoder to generate the image of the cross-sectional slice i

[0091] In a specific embodiment, I s , I c It is the image of the central slice. This image is the result of the combined effect of all thermal factors in the furnace. It can be considered that this image contains some information about the entire temperature field. In order to extract this information, the feature encoder uses resnet for feature extraction. Resnet extracts color information at a shallow level and structural information at a deep level. Since the temperature information in the temperature field requires not only structural information but also color information, the information extracted from the encoder is the feature of the middle level. Resnet is used to extract I s , I c Feature extraction is performed separately to obtain:

[0092] Res(I s , I c )=(f s , f c )

[0093] Among them, f s , f c Indicates that from I s , I c Extracted intermediate level feature information.

[0094] In a specific embodiment, a set of standard 3D positions, 128×128×128 spatial sampling points p are defined.g The corresponding value of each dimension position is from 0 to 127. The method designed by the present invention can be applied to any N×N×N if the computing resources allow. In the case of other resolutions, the 3D position also needs to be modified accordingly. After the 3D position is generated, a three-dimensional structure composed of 128×128×128 spatial voxels is formed (standard). The observation I is defined s Position V s and Observation I c Position V c With the above conditions, ray casting can be performed. In the present invention, I s , I c Each pixel value is regarded as a ray of light projected through the furnace 3D space volume to I s , I c The projection value of the surface where it is located, this process just simulates the combined action of all thermal factors in the furnace to generate I s , I c The effect of this is that each ray passes through I s , I c The intermediate level features are resampled to the 128×128×128 spatial volume to reconstruct each voxel in the furnace space. This embodiment adopts a slice-by-slice generation method. For each 3D point p g_i =(x g_i ,y g_i ,z g_i )To I s , I c projection:

[0095] (u i ,v i ) T =[R|t](x g_i ,y g_i ,z g_i ) T ;

[0096] Where R and t represent the rotation matrix and translation vector respectively. s , I c The extracted intermediate level features f s , f c In association, the method B of sampling bilinear interpolation of the association method is expressed as:

[0097] f i =C(B(f s ,u s_i ,v s_i ),B(f c ,u c_i ,v c_i ))

[0098] Where C represents the stacking operation, f i That is to reconstruct the features of slice i, (u s_i ,v s_i ) is the sagittal center slice projection point, (u c_i ,v c_i ) is the projection point of the coronal center slice.

[0099] In a specific embodiment, the decoder part of the diffusion model is used to decode the features of the reconstructed slice i to obtain the final image, which is expressed as:

[0100]

[0101] in, is the image of the cross-section slice i, and D is the decoder.

[0102] In this embodiment, the feature encoder, feature space projection, and feature decoder form the generator part, and the discriminator part uses the pix2pix discriminator. The loss function of the generator is:

[0103] L=w1L1+w2L2+w3L3

[0104] L1 is rebuilt The cross entropy loss between the true value and L2 is the reconstruction The perceptual loss between the L3 and the true value is expressed as

[0105]

[0106] in, Indicates the pixel value of the middle column of generated slice i, I s (i,:) represents I s The pixel value of the i-th row, Indicates the pixel value of the middle row of generated slice i, I c (i,:) represents I c The pixel value of the i-th row of , w1, w2, w3 are coefficients, set to 1, 10, 1 respectively. The discriminator loss function is the same as pix2pix.

[0107] In a specific embodiment, further subsequent processing is included, and the process includes:

[0108] Figure 7The output image of the network shows the result. Since the boiler temperature field is irregular, in order to unify the resolution of the generated slices, white edges are added to the upper and lower parts of the generated temperature map. However, since the bilinear interpolation method is embedded in the network, abnormal temperature values ​​will be generated at the edges of the temperature map and the white edge area, resulting in a lot of noise points. A post-processing process is added to define a mask template for each generated slice, such as Figure 8 , where white is defined as valid data 1 and black is defined as invalid data 0. These noise points can be removed by multiplying the mask with the corresponding elements of the result image, such as Fig. 9 .

[0109] Experimental analysis of the first phase: The experimental analysis of the first phase includes two parts. The first part is the fitting effect of the multi-layer perceptron, and the second part is the effect of the coarse-to-fine inverse distance weighted interpolation method.

[0110] Analysis of the experimental results of multi-layer perceptron fitting: The multi-layer perceptron fitting is the direct temperature measurement point and the water-cooled wall temperature measurement point to I s , I c The temperature change of the upper projection point is not directly fitted in this embodiment, but with the help of the reference working condition, the input data of the multi-layer perceptron fitting becomes the difference ΔT and ΔT between the direct temperature measurement point and the water-cooled wall temperature measurement point between the current working condition and the reference working condition. w At the same time, the multi-layer perceptron designed by the present invention calculates ΔT and ΔT w The projections in the coronal and sagittal planes are fitted separately. This design can greatly improve the fitting accuracy. The corresponding accuracy evaluation indicators for F1, G2, and F2 are shown in Table 2. The evaluation indicators are root mean square error and R 2 The values ​​are evaluated. From the data in the table, it can be seen that the accuracy of the fitted data is high.

[0111] Table 3 compares the method of the present invention, F1, G2, F2 without separate fitting and input data without difference ΔT, ΔT w The method where the input data is not C. r The water-cooled wall temperature and the temperature of the direct temperature measurement point are not separated. F1, G2, and F2 are not used for separate fitting. It means that a multi-layer perceptron model is used to fit the projected temperature value. The evaluation indicators are the root mean square error and R 2 It can be seen from the table that the calculation accuracy of the method of the present invention has been significantly improved. This is because separate fitting can learn different nonlinear relationships through multiple groups of parameters. Since the physical principles of temperature values ​​generated by direct temperature measurement points and water-cooled wall temperature measurement points are different, different nonlinear relationships are used for fitting better results. At the same time, I s , I cDifferent projection directions should also use different nonlinear relationships, so the effect of separate fitting of multiple multi-layer perceptrons used in the present invention is better than that of not using separate fitting. In addition, the present invention selects the difference ΔT and ΔT w As input, it can eliminate the influence of various random factors between different groups of data and reduce the impact of uncertainty factors in the boiler on temperature distribution.

[0112] Table 2 F1, G2, F2 accuracy evaluation

[0113]

[0114] Table 3 The data of the present invention do not use the difference method to evaluate the accuracy

[0115]

[0116]

[0117] The experimental effect analysis of the inverse distance weighted interpolation method: The present invention adopts a multi-layer inverse distance weighted interpolation method to obtain the reference working condition and the current working condition in I s , I c Difference map I d_s , I d_c , and then calculate I s , I c ,The figure shows the calculation accuracy of the multi-layer IDW ,interpolation method (using the difference map), the single layer IDW ,interpolation method and the multi-layer IDW ,interpolation method (not using the difference map). The accuracy indicators are compared using PSNR, SSIM, and MAE indicators, as shown in Table 4.

[0118] Table 4 Comparison of accuracy indicators

[0119] coronal plane Sagittal plane LPIPS maximum 0.1087 0.1160 LPIPS minimum 0.0479 0.0227 LPIPS average 0.0713 0.0658 Maximum PSNR 25.7281 28.4632 PSNR minimum 19.7064 18.9022 Average PSNR 23.5508 24.2768 SSIM maximum 0.8598 0.9174 SSIM minimum 0.7895 0.8103 SSIM average 0.8268 0.8611

[0120] The generative adversarial neural network generates a slice-by-slice temperature field, and the results are as follows Fig.10 shown.

[0121] A furnace 3D temperature field reconstruction system, comprising:

[0122] Database establishment module, establishes a reference database of temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images;

[0123] The slice image calculation module obtains the new working condition data, selects the working condition data most similar to the new working condition data based on the reference database as the reference working condition data, obtains the projected temperature points and the corresponding temperature change values ​​according to the reference working condition data and the new working condition data; obtains the temperature difference map according to the projected temperature points and the corresponding temperature change values ​​combined with the inverse distance weighted interpolation method, and calculates the sagittal and coronal central slice images of the new working condition data based on the temperature difference map;

[0124] The temperature field reconstruction module uses the sagittal and coronal central slice images of the new operating condition data to simulate the combined effect of thermodynamic factors in the furnace through ray projection, and combines the pre-trained adversarial neural network to reconstruct all cross-sectional images of the new operating condition data to obtain a reconstructed 3D furnace temperature field.

[0125] The present invention proposes a method for quickly and accurately reconstructing a complete N×N×N temperature field, fully considering the correction effect of the water-cooled wall temperature measurement point on the temperature field, so that the generated coronal and sagittal center slice images are closer to the actual temperature distribution in the furnace, and creatively proposes a method for mapping 2D features to 3D space through a ray projection model, which can restore the entire three-dimensional temperature field from the coronal and sagittal center slice images. This method shows great potential in practical applications, not only can it ensure the accuracy of the generated temperature field, but also can greatly improve the efficiency of temperature field reconstruction and ensure the optimization effect of the furnace thermal management strategy.

[0126] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0127] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for reconstructing a 3D temperature field in a furnace, characterized in that: include: Establishing a reference database of the temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images; Acquire new working condition data, select working condition data that is most similar to the new working condition data based on the reference database as reference working condition data, acquire projected temperature points and corresponding temperature change values ​​based on the reference working condition data and the new working condition data; acquire a temperature difference map based on the projected temperature points and the corresponding temperature change values ​​in combination with an inverse distance weighted interpolation method, and calculate sagittal and coronal central slice images of the new working condition data based on the temperature difference map; Through the sagittal and coronal central slice images of the new operating condition data, the combined effect of thermodynamic factors in the furnace is simulated by ray projection, and all cross-sectional images of the new operating condition data are reconstructed in combination with the pre-trained adversarial neural network to obtain the reconstructed 3D furnace temperature field.

2. A method for reconstructing a 3D temperature field in a furnace according to claim 1, characterized in that: The selection of reference operating condition data that is most similar to the new operating condition data based on the reference database specifically includes: finding the load corresponding interval of the new operating condition data, and finding the operating condition that is most similar to the new operating condition data and the operating condition data in the interval through KNN clustering as the reference operating condition data; the KNN clustering features select the direct temperature measurement point data and boundary condition data.

3. The method for reconstructing a 3D temperature field in a furnace according to claim 1, characterized in that: The obtaining of the projected temperature points and the corresponding temperature change values ​​according to the reference operating condition data and the new operating condition data specifically includes: the projected temperature points include water-cooled wall temperature points and direct temperature measurement points; is the central slice image I in the sagittal position s The water-cooled wall temperature point projected on the top, The central slice image I in the coronal position c The temperature difference between the new working condition data and the reference working condition data at the water-cooled wall temperature point is ΔT. The projected temperature value should have a corresponding change. The change satisfies the nonlinear transformation relationship. The formula is as follows: Where, ΔT c Indicates that the direct temperature measurement point is in the central slice image I of the coronal position c Temperature change value of the upper projection point, ΔT w_s ,ΔT w_c Respectively represent the water-cooled wall temperature measurement point in the sagittal and coronal center slice images I s , I c The temperature change value of the upper projection point, F1, G2, F2 represents the nonlinear transformation relationship.

4. A method for reconstructing a 3D temperature field in a furnace according to claim 3, characterized in that: The method of calculating the sagittal and coronal central slice images of the new working condition data based on the temperature difference map specifically includes: using the inverse distance weighted interpolation method to propagate the temperature change value to I s , I c Get the temperature difference map on the whole image; The sparse projection set {ΔT p |ΔT c ,ΔT w_s ,ΔT w_c } as the observed value, I s , I c The coordinate points to be filled in are regarded as points to be estimated, and the temperatures to be estimated are regarded as values ​​to be estimated; Calculate the weight coefficients of the M observation points closest to the point to be estimated for the point to be estimated, as follows: Among them, W idw is the weight coefficient, (u0,v0) is the spatial coordinate of the point to be estimated, (u j ,v j ) is the spatial coordinate of the observation point, d j is the Euclidean distance from the estimated point j to the observation point, is the value to be estimated, q is a constant, T j are the j closest observation points; The multi-layer inverse distance weighted interpolation method is used to interpolate the projection points to obtain the intermediate point set, and then interpolate the intermediate point set again to obtain I s , I c All the points in the above figure are interpolated to obtain the temperature difference between the sagittal and coronal central slices. d_s , I d_c , combined with the sagittal and coronal central slice images corresponding to the reference working condition data I r_s , I r_c , get I s , I c , as shown below: I s ,I c =I r_s ,I r_c +I d_s ,I d_c 。 5. A method for reconstructing a 3D temperature field in a furnace according to claim 4, characterized in that: Reconstructing all the cross-sectional images of the new working condition data specifically includes: s , I c Perform feature extraction, extract the position of slice i from the 3D position, and convert the feature information, slice position, and sagittal observation position V s and coronal observation position V c Perform ray projection, combine the feature information with the 3D position through feature space projection, and then decode the feature through the decoder to generate the image of the cross-sectional slice i 6. A method for reconstructing a 3D temperature field in a furnace according to claim 5, characterized in that: The encoder is used to s , I c The feature extraction specifically includes: the feature encoder uses resnet to extract features, and uses the resnet to extract I s , I c Feature extraction is performed separately to obtain: Res(I s ,I c )=(f s ,f c ) Among them, f s , f c Indicates that from I s , I c Extracted intermediate level feature information.

7. The method for reconstructing a 3D temperature field in a furnace according to claim 5, characterized in that: The combining of the feature information with the 3D position through feature space projection specifically includes: for a slice i in the 3D position, for each 3D point p g_i =(x g_i ,y g_i ,z g_i )To I s , I c projection: (u i ,v i ) T =[R|t](x g_i ,y g_i ,z g_i ) T ; Where R and t represent the rotation matrix and translation vector respectively. s , I c The extracted intermediate level feature information f s , f c In association, the method B of sampling bilinear interpolation of the association method is expressed as: f i =C(B(f s ,u s_i ,v s_i ),B(f c ,u c_i ,v c_i )) Where C represents the stacking operation, f i That is to reconstruct the features of slice i, (u s_i ,v s_i ) is the sagittal center slice projection point, (u c_i ,v c_i ) is the projection point of the coronal center slice.

8. A method for reconstructing a 3D temperature field in a furnace according to claim 7, characterized in that: The feature decoding implemented by the decoder specifically includes: in, is the image of the cross-section slice i, and D is the decoder.

9. A furnace 3D temperature field reconstruction system, using a furnace 3D temperature field reconstruction method according to any one of claims 1 to 8, characterized in that: include: A database establishment module is used to establish a reference database of the temperature field; the reference database includes direct temperature measurement point data, water-cooled wall temperature measurement point data, boundary condition data, sagittal and coronal center slice images, and all transverse slice images; a slice image calculation module, which obtains new working condition data, selects working condition data that is most similar to the new working condition data based on the reference database as reference working condition data, obtains projected temperature points and corresponding temperature change values ​​based on the reference working condition data and the new working condition data; obtains a temperature difference map based on the projected temperature points and the corresponding temperature change values ​​in combination with an inverse distance weighted interpolation method, and calculates sagittal and coronal central slice images of the new working condition data based on the temperature difference map; The temperature field reconstruction module uses the sagittal and coronal central slice images of the new operating condition data to simulate the combined effect of thermodynamic factors in the furnace through ray projection, and combines the pre-trained adversarial neural network to reconstruct all cross-sectional images of the new operating condition data to obtain a reconstructed 3D furnace temperature field.

Images